[AgTech Rising] AI Sprays Only Weeds… GINT to Commercialize Upland-Field Weeding Robots in 2029

According to the Rural Development Administration’s “2023 Agricultural Work Mechanization Rate Statistics,” the mechanization rate for paddy rice was 99.7%, while that for upland crops remained at only 67.0%. Sowing and transplanting (18.2%) and harvesting (42.9%) are especially low. The agricultural workforce is getting older by the day. In the National Data Agency’s “2025 Census of Agriculture, Forestry and Fisheries,” 51.3% of the farm population was aged 65 or older, 2.5 times the overall population aging rate (20.3%). Even so, weeding in upland fields still relies on people finding weeds one by one and spraying chemicals. 

A national R&D project to convert this process to unmanned robots began in 2026. It is the “Development of Unmanned Collaborative Precision Weeding Technology Based on Data Solutions” project, led by agricultural robotics company GINT Co., Ltd. As a collaborative research project jointly promoted by the Ministry of Agriculture, Food and Rural Affairs and the Rural Development Administration, it will run for three years and nine months, from April 2026 to December 2029. Total research funding is KRW 6.15 billion.

The project is divided into two tracks. The Agriculture Ministry project (KRW 3.75 billion) is led by GINT and is responsible for the driving platform, which corresponds to the robot’s body, and the precision spraying system. The Rural Development Administration project (KRW 2.4 billion) is led by Sunchon National University and will develop the AI recognition model and data solution that distinguish crops from weeds—in other words, the robot’s “brain.”
The GINT consortium includes Semyung Soft, the Korea Textile Machinery Convergence Research Institute, Jeonbuk National University, Pusan National University, and the Jeonbuk State Agricultural Research and Extension Services, while the Rural Development Administration consortium is joined by the Jeonnam-Gwangju Integrated Special City Agricultural Research and Extension Services. The target performance is a weed recognition rate of at least 95%, a weed removal success rate of at least 90%, and an improvement in collaborative work efficiency of at least 45%.
 

In targeted-control simulation, pesticide use per ha falls from 200ℓ to 120ℓ   

GINT Chief Researcher Jeong Hee-jong, who oversees the project, cited the structure of Korean upland fields as the reason why mechanization of upland-field weeding has lagged. Korean upland fields are cultivated by making ridges and furrows, but their dimensions differ by region and by crop, meaning there is still no common equipment that can be used broadly. “At present, what can be called mechanization in upland farming is limited to attaching implements to ride-on cultivators,” Jeong said. “Because the equipment is so large, crops are often damaged when it turns at the end after driving straight, and for small and medium-sized farms, mechanization itself is difficult to apply.”

The solution GINT envisions is “targeted control,” which sprays chemicals only on selected weeds. At present, even when spraying is done by machine, chemicals must be sprayed over the entire field. To spray only selected weeds, it has to be done by hand. The explanation is that if a robot identifies weeds and sprays a fixed amount only at the relevant points, it can greatly reduce chemical use while making the work unmanned. In GINT’s simulation, which assumes certain conditions such as the area where weeds occur, the amount of chemical used under conventional control—about 200ℓ per ha—falls to around 120ℓ. GINT plans to verify and refine this figure through prototype testing.

The robot is electric. The technology development goal is to operate for at least six hours on a single charge and work about 2 ha. GINT explains that, in concept, 24-hour operation is also possible if the battery is replaced. 
 

Multiple small robots instead of one large unit… Collaboration means division of labor

A distinctive feature of this project is that it is structured around multiple small robots working together, rather than a single large robot. “Unlike Europe or the United States, Korea has a diverse distribution of small and medium-sized farms, so miniaturized equipment is needed,” Jeong said. “At the same time, so that it can also be used on large farmland, two or more units share work information and routes and divide the work. Collaboration can be understood as division of labor.”

This structure does not stop at weeding. GINT’s long-term goal is to bundle the entire upland-farming process—from ridge shaping, sowing, weeding, pest and weed control, and management—into a single system. To this end, it is monitoring all machines currently being sold through its own service server. Based on this technology, it also plans to commercialize a control system that manages multiple robots deployed in one work space.

Asked whether overseas products can be used as they are in Korean upland fields, Jeong replied, “The specifications themselves simply do not match.” In Europe and elsewhere, crops are cultivated on flat land without ridges and furrows, but in Korea, ridges are built and crops are planted densely, making the farming method itself different. He explained that to use overseas robots, a separate farming method tailored to them would have to be developed. GINT aims to respond to the varied specifications of Korean upland fields with active track-width (distance between left and right wheels) and ground-clearance adjustment technology, while also setting the price at 30–35% of advanced overseas products.
 

Soybeans at the cotyledon stage are distinguished from weeds using sowing-interval data and images

Distinguishing young crops from weeds is the core challenge for a weeding robot. At the cotyledon stage, soybeans are difficult to distinguish from weeds using images alone. GINT uses the characteristics of soybean sowing, which has already been partly mechanized. “If there is data on the row spacing of planted crops, the crop positions can be predicted,” Jeong said. “If the predicted positions match the image information recognized by the trained model, they are judged to be soybeans.” According to Jeong, once true leaves begin to emerge, soybeans and weeds can be completely separated using images alone. 

The 10% missed in one round of control is made up through repeated control. Since weeding is not a task that ends in a single pass, the judgment is that two to three rounds of control can produce an effect close to virtually 100%. In furrows where there are no crops and only weeds grow thickly, a combined approach using mechanical weeding is applied. For chemicals, farmers use the existing herbicides or organic farming materials they had been using as they are.

The project is currently in the data-gathering stage. At the test field of the Jeonbuk State Agricultural Research and Extension Services, paddy soybeans and mung beans are being studied, while at the Jeonnam-Gwangju Integrated Special City Agricultural Research and Extension Services, Jeonbuk National University and Sunchon National University are each cooperating to collect crop and weed image data on local specialty crops. A “camera driving platform for data collection,” which reproduces the environment in which the weeding robot recognizes crops, moves around fields and records how crops grow and weeds spread by season. Using these images, the Rural Development Administration consortium trains AI and applies it to spray-nozzle control. The first prototype is scheduled to be completed by June 2027 and deployed in actual fields.
 

Rental centers and individual sales in parallel… Standards and testing criteria for unmanned weeding robots remain a task

Multiple distribution methods are being left open. Jeong said that because agricultural machinery rental centers rent equipment to an unspecified number of users, robots could be distributed there in a form equipped only with the weeding function and without data updates. By contrast, he expects that models that collect data in real time, learn, update, and improve performance will center on individual sales. “The structure allows a spraying unit or other implements to be swapped onto the driving platform, so once the platform is purchased at first, only the implements need to be bought afterward,” he said. “Consumers’ purchase burden will decrease, and if policy support is also provided, there should be no major problem with marketability.”

GINT does not see this robot only as agricultural weeding equipment. It is approaching the market as an “outdoor service robot” that expands its uses into outdoor spaces in other industries, such as golf courses or park green areas that require control work.

Safety standards and standardization are tasks to be resolved going forward. There are still no domestic dedicated standards or testing criteria for unmanned weeding robots. GINT plans to complete demonstration of unit robots by 2027 and propose standards around 2028, when demonstration data on multi-robot collaboration will have accumulated in the second-stage project. According to Jeong, under current law there is still no basis for unmanned autonomous agricultural machinery to drive on public roads. Accordingly, autonomous driving is restricted outside farmland, and in a separate project GINT is developing an AI model that drives at low speed on farm roads while recognizing and avoiding people, vehicles, tractors, waterways, and other obstacles. This technology will also be applied to the weeding robot at the commercialization stage.

Regarding what rural areas will look like in 2029, when the project ends, Jeong said, “We are envisioning a scene in which farmers control the machines from inside a room, and unmanned robots move through actual fields.”